The pervasive effects of argument length on inductive reasoning

Comments

Transcription

The pervasive effects of argument length on inductive reasoning
THINKING & REASONING, 2012, 18 (3), 244–277
The pervasive effects of argument length on inductive
reasoning
Downloaded by [University of California Merced] at 08:37 11 June 2012
Evan Heit1 and Caren M. Rotello2
1
School of Social Sciences, Humanities and Arts, University of California,
Merced, Merced, CA, USA
2
Department of Psychology, University of Massachusetts Amherst,
Amherst, MA, USA
Three experiments examined the influence of argument length on plausibility
judgements, in a category-based induction task. The general results were that
when arguments were logically invalid they were considered stronger when
they were longer, but for logically valid arguments longer arguments were
considered weaker. In Experiments 1a and 1b when participants were
forewarned to avoid using length as a cue to judging plausibility, they still
did so. Indeed, participants given the opposite instructions did not follow
those instructions either. In Experiment 2 arguments came from a reliable or
unreliable speaker. This manipulation affected accuracy as well as response
bias, but the effects of argument length for both reliable and unreliable
speakers replicated Experiments 1a and 1b. The results were analysed using
receiver operating characteristic (ROC) curves and modelled using multidimensional signal detection theory (SDT). Implications for models of
category-based inductive reasoning, and theories of reasoning more generally,
are discussed.
Keywords: Inductive reasoning; Argumentation; Mathematical modelling.
Correspondence should be addressed to Evan Heit, School of Social Sciences, Humanities
and Arts, University of California, Merced, 5200 North Lake Road, Merced, CA 95343, USA.
E-mail: [email protected]
This work was supported by National Science Foundation collaborative research grant
BCS-0616979. We thank Lissette Alvarez, Melony Bowling, Brooklynn Edwards, Efferman
Ezell, Chanita Intawan, Jascha Ephraim, Markie Johnson, Alex Parnell, Nic Raboy, Haruka
Swendsen, and Jonathan Vickrey for assistance with this research.
Ó 2012 Psychology Press, an imprint of the Taylor & Francis Group, an Informa business
http://www.psypress.com/tar
http://dx.doi.org/10.1080/13546783.2012.695161
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
245
The expression ‘‘weighing the evidence’’ suggests that making a
judgement depends on assessing a quantity or mass. Whether or not an
argument seems plausible depends on the amount of evidence presented in
its favour. Presumably arguments with more evidence in the premises are
stronger arguments. Research has shown that people are highly influenced
by the length of an argument, but the situation is more complicated. One
argument might be very long but still invalid, and another argument might
be very short and perfectly valid. It’s also possible to imagine that, in some
cases, short parsimonious arguments would be more convincing than long
rambling arguments.
In this paper we address the pervasive effects of argument length on
judgements of plausibility. First we review previous research on social
cognition, argumentation, category-based inductive reasoning, and causal
reasoning on the effects of argument length, as well as our own research
(Rotello & Heit, 2009). Then we review theoretical accounts of reasoning
that make predictions about the effects of argument length on plausibility
judgements. Finally we present three new experiments that examine whether
people can avoid using argument length when judging plausibility, or if
argument length is so compelling that people cannot help being influenced
by it.
PREVIOUS RESEARCH ON ARGUMENT LENGTH
Social cognition
Classic research on attitude change and persuasion has identified number of
arguments as a key variable affecting belief in an overall conclusion. For
example, Cook (1969) found that people were more likely to agree with
counter-attitudinal statements like ‘‘cleaning the teeth more than once per
week is harmful to health’’ when more supporting arguments were presented
in its favour. Burnstein, Vinokur, and Trope (1973) examined the
phenomenon of group polarisation in a risky choice situation. They found
that the number of arguments presented in favour of a risky shift, rather
than the number of people making arguments, predicted attitude change. In
a mock jury study Calder, Insko, and Yandell (1974) varied the number of
prosecution and defence arguments, and found that jurors were more likely
to convict when there were more prosecution arguments and when there
were fewer defence arguments. Petty and Cacioppo (1984) asked college
students whether they supported a somewhat unappealing proposal to
introduce comprehensive exams for graduating seniors. This proposal was
presented along with three or nine arguments, which themselves were either
of high or low quality. When this was a low-involvement issue (the exams
would affect students 10 years in the future) students were not influenced by
246
HEIT AND ROTELLO
Downloaded by [University of California Merced] at 08:37 11 June 2012
the quality of the arguments, and responded simply based on the number of
arguments. However, when this was a high-involvement issue (current
students would be affected) argument quality affected attitudes. Moreover
there was an interaction between number of arguments and quality:
Additional strong arguments increased support for the proposal, but
additional weak arguments reduced support. Thus the Petty and Cacioppo
study provided important boundary conditions on the notion that more
evidence is better.
Argumentation
Closely related to the social cognition literature is research on argumentation, which has sought to identify structural characteristics that make
arguments stronger or weaker. O’Keefe (1997, 1998) conducted metaanalyses to examine the effects of making arguments (or standpoints) more
explicit (and hence longer). One could imagine that more explicit arguments
would be more convincing because they provide more evidence, but on the
other hand more explicit arguments may be less engaging, more open to
scrutiny, and more likely to provoke disagreement. O’Keefe found that
spelling out conclusions in detail, identifying sources, making implicit
premises explicit, and providing precise quantitative information all made
arguments seem stronger. Relatedly, O’Keefe (1999) conducted a metaanalysis to compare one-sided arguments, in which a positive argument is
made for a position, to two-sided arguments, in which opposing views are
acknowledge or even refuted. Two-sided arguments will naturally tend to be
longer than one-sided arguments, but they run the risk of making counterarguments more salient. O’Keefe found that two-sided arguments were seen
as more convincing than one-sided arguments, and refutational two-sided
arguments were seen as more convincing than nonrefutational two-sided
arguments. Hence the general trend in the argumentation literature points to
longer arguments being stronger (although note that merely adding
anecdotal evidence to increase length does not necessarily increase strength,
e.g., Hornikx & Hoeken, 2007).
Category-based induction
Research on inductive reasoning has examined the effects of varying the
number of premises within a single argument (for a review see Heit, 2000).
Nisbett, Krantz, Jepson, and Kunda (1983) systematically varied the given
number of observations in an estimation task. For example, participants
were told that 1, 3, or 20 obese members of the Barratos group had been
observed, and were asked what proportion of all Barratos are obese. In
general, inferences were stronger with increased sample size. (Strictly
ARGUMENT LENGTH AND INDUCTIVE REASONING
247
Downloaded by [University of California Merced] at 08:37 11 June 2012
speaking this is a sample size manipulation rather than an argument length
manipulation, but our aim here is to be as inclusive as possible.)
Osherson, Smith, Wilkie, Lopez, and Shafir (1990) identified a
phenomenon called monotonicity in which arguments with more premises
are considered more plausible, for example, (2) is considered a stronger
argument than (1).
Sparrows have sesamoid bones
Eagles have sesamoid bones
——————————————
All birds have sesamoid bones
(1)
Hawks have sesamoid bones
Sparrows have sesamoid bones
Eagles have sesamoid bones
——————————————
All birds have sesamoid bones
(2)
This monotonicity effect appears to be very robust: McDonald, Samuels,
and Rispoli (1996) reported that, over a large set of arguments, there was a
positive correlation between number of premises and judged argument
strength. It also appears that the monotonicity effect is reasonably general:
Lopez, Gelman, Gutheil, and Smith (1992) and Gutheil and Gelman (1997)
reported some evidence for monotonicity effects in 9-year-old children.
Interestingly, in a study of individual differences in adults, Feeney (2007)
found that adult participants with higher intelligence showed greater
monotonicity effects.
Note, however, that having more premises does not always lead to
stronger arguments. Osherson et al. (1990) documented exceptions to the
monotonicity effect, called non-monotonicity effects, in which longer
arguments seem weaker, as in (3) and (4).
Crows secrete uric acid crystals
Peacocks secrete uric acid crystals
———————————————
All birds secrete uric acid crystals
(3)
Crows secrete uric acid crystals
Peacocks secrete uric acid crystals
Rabbits secrete uric acid crystals
——————————————––
All birds secrete uric acid crystals
(4)
Here, adding premise information about rabbits that is not obviously
relevant to the conclusion weakens the argument. In this case rabbits fall
into a different superordinate category than the birds that are the focus
of the other premises in the argument. However, Sloman (1993) showed
248
HEIT AND ROTELLO
Downloaded by [University of California Merced] at 08:37 11 June 2012
non-monotonicity effects even within a single superordinate category, for
example, (6) was considered weaker than (5).
All crocodiles have acidic saliva
——————————————–
All alligators have acidic saliva
(5)
All crocodiles have acidic saliva
All king snakes have acidic saliva
——————————————–
All alligators have acidic saliva
(6)
Here all of the animals are reptiles, but king snakes are so dissimilar to
alligators that adding this second premise seems to weaken the argument.
Causal reasoning
In category-based induction tasks participants evaluate a conclusion based
on a premise or set of premises. In causal reasoning tasks the task is often
reversed: participants are given a conclusion and asked to evaluate a premise
or set of premises in terms of how well they explain that conclusion.
Borrowing an example from Read and Marcus-Newhall (1993), imagine
that Cheryl is tired, is frequently nauseous, and is gaining weight—
collectively, these observations are a conclusion to be explained. The shorter
explanation that Cheryl is pregnant seems more compelling than the longer
explanation that Cheryl has mononucleosis, has a stomach virus, and has
stopped exercising. The idea of preferring shorter or simpler explanations
follows from the principle of Occam’s razor that ‘‘entities should not be
multiplied unnecessarily’’ (for a review see Lombrozo, 2007) and it also
follows from Thagard’s (1989) connectionist network model of explanatory
coherence. Both Read and Marcus-Newhall, and Lombrozo, provided
experimental evidence that people favour shorter and simpler explanations
over longer and more complicated explanations, with Lombrozo showing
particularly strong results by controlling for prior probability of various
explanations.
Whereas most findings on social cognition, argumentation, and categorybased induction indicate that longer arguments are more convincing than
shorter arguments, in research on causal reasoning it has been found that
shorter explanations are more convincing than longer explanations.
However, even in the social cognition and category-based induction
literatures, there are some exceptions to the generalisation that longer
arguments are more convincing. So the picture is somewhat unclear on
whether longer arguments are better or worse than short arguments. Of
course all of these studies varied in numerous ways, making it difficult to
ARGUMENT LENGTH AND INDUCTIVE REASONING
249
draw a general conclusion about argument length. We next turn to our own
research on argument length, which has found both positive and negative
effects of argument length within a single experimental paradigm.
Downloaded by [University of California Merced] at 08:37 11 June 2012
Rotello and Heit (2009)
In a recent study of category-based induction we manipulated the length of
arguments while also manipulating whether they are logically valid.
(Although an argument encompasses both premises and a conclusion, we
varied length just in terms of number of premises.) The arguments had one,
three, or five premises, and participants judged the plausibility of the
conclusion. An example invalid argument, with three premises, is shown in
(7).
Horses have Property X
Mice have Property X
Sheep have Property X
——————————–
Cows have Property X
(7)
An example valid argument, with five premises, is shown in (8).
Horses have Property X
Mice have Property X
Sheep have Property X
Rabbits have Property X
Cats have Property X
———————————
Rabbits have Property X
(8)
For invalid arguments we found the usual monotonicity effect in
category-based induction. Five-premise arguments were stronger than
three-premise arguments, which in turn were stronger than one-premise
arguments. However, the reverse pattern was found for valid arguments.
One-premise arguments were the strongest and five-premise arguments were
the weakest. This reverse pattern for valid arguments resembles the results
from research on causal explanation. If the premises in an argument are
conceived of an explanation for the conclusion, then once there is a sufficient
explanation for the conclusion, adding additional premises seems to weaken
the argument overall.
Although Rotello and Heit (2009) found consistent effects of argument
length on judgements of plausibility, we also identified a related task in
which participants showed little influence of argument length. When
participants were asked to judge logical validity rather than plausibility
they showed a greater influence of validity itself and reduced influence of
Downloaded by [University of California Merced] at 08:37 11 June 2012
250
HEIT AND ROTELLO
length. (Although length did not have a statistically significant influence on
validity judgements, the trends were the same for validity as for plausibility,
so it is impossible to rule out a small effect of length on validity judgements.)
We explained this finding in terms of a two-dimensional account of
reasoning that was implemented as a multidimensional signal detection
model and fitted successfully to the data. One dimension corresponded to
sensitivity to the rules of logic, and the other dimension used associative
information such as argument length. The difference between the plausibility
and validity judgements was accounted for in terms of different relative
impacts of the two dimensions on the two kinds of judgement: Plausibility
judgements were influenced about equally by the two dimensions, whereas
validity judgements were influenced more by the logic-based process. The
model is illustrated in Figure 1, where the key difference between plausibility
and validity judgements is the slope of the decision bound.
THEORETICAL ACCOUNTS
The two-dimensional nature of the Rotello and Heit (2009) model of
reasoning (see also Heit & Rotello, 2010; Heit, Rotello, & Hayes, 2012) is in
accord with Petty and Cacioppo’s (1984) own two-process explanation,
which referred to central and peripheral processing of information, and
Figure 1. Schematic of Rotello and Heit’s (2009) model of inductive and deductive reasoning.
The same evidence distributions (ellipses) are used for both types of judgement, but different
decision bounds (shown as dashed lines). Three distributions are shown for invalid arguments;
those to the right reflect invalid arguments with more premises. Two valid distributions are
shown; those to the left reflect valid arguments with more premises. Reprinted with permission.
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
251
itself is representative of theoretical work in social psychology. In addition,
two-process accounts of reasoning have been very influential in cognitive
psychology (Evans, 2008; Sloman, 1996; Stanovich, 2009). What all these
accounts have in common is the notion that some processing, for example
noticing the length of an argument, is automatic. Potentially people could
avoid using this information—this is what Evans calls an ‘‘intervention’’ and
Stanovich calls an ‘‘override’’. However, what is then required is some other
process to substitute for the automatic process. In the Rotello and Heit
study some participants were explicitly instructed to pay attention to logical
validity, and these participants were indeed able to substitute logical
processing for processing based on more superficial information such as
argument length. In the present experiments we focused on plausibility
judgements, and whether people could avoid the use of argument length
when judging plausibility. Under the assumption that noticing the length of
an argument is automatic, and without explicit instructions to make
judgements on some other basis, we predicted that it would be very difficult
to avoid using argument length in judgements of plausibility. Put another
way, argument length is intrinsic to judgements of plausibility.
Indeed, theoretical accounts of inductive reasoning are necessarily
affected by length of argument (for reviews see Hayes, Heit, & Swendsen,
2010; Heit, 2008). Osherson et al. (1990) presented a model that accounted
for plausibility judgements in terms of the maximum similarity between
categories in the premises of an argument and the category in the conclusion
of an argument (as well as members of the superordinate category which
includes all of the other categories). Adding premises will generally increase
the maximum level of similarity. For example, in going from (1) to (2),
adding the premise about hawks will increase the similarity to hawk-like
members of the bird category, in effect increasing the coverage of this
category. The main exception is for non-monotonicity effects as in (4)—
when the premise about rabbits is introduced, the superordinate category
under consideration is all animals rather than birds, as in (3). The larger
category is harder to cover. Hence coverage decreases from (3) to (4).
However, in our own experiments the superordinate category is always
mammal, so this exception does not apply.
Sloman’s (1993) model of inductive reasoning also leans on the
concept of coverage, which is implemented via feature overlap in a
connectionist network. The network is trained on each successive premise,
and additional premises can only increase the level of activation in the
network and the potential overlap with the conclusion. This model
predicts that making an argument longer will always make its conclusion
more plausible.
Finally, Bayesian models of inductive reasoning (Heit, 1998, 2000; Kemp
& Tenenbaum, 2009; Tenenbaum & Griffiths, 2001) generally predict that
Downloaded by [University of California Merced] at 08:37 11 June 2012
252
HEIT AND ROTELLO
longer arguments will have their conclusions judged more plausible. These
models operate in terms of a space of competing hypotheses. For example,
in (1) and (2) the hypothesis that all birds have sesamoid bones is competing
with other hypotheses, for example the hypothesis that only sparrows and
eagles have sesamoid bones. Adding the premise that hawks have sesamoid
bones rules out the latter hypothesis, so there is less competition for the
hypothesis that all birds have sesamoid bones. In general, adding premises
will rule out some prior hypotheses and strengthen the remaining ones. In
Kemp and Tenenbaum’s terms, when adding premises some hypotheses
from the prior distribution are no longer possible; in the posterior
distribution for the hypothesis space probabilities are renormalised and
the remaining hypotheses increase in probability. For Bayesian models the
main exception would be if what Tenenbaum and Griffiths called a size
principle is incorporated. Under this principle there is a bias to favour more
restricted conclusions over broader conclusions, and this bias is increasingly
manifest as more observations are made. The size principle could account
for the non-monotonicity effects reported by Osherson et al. (1990) and
Sloman (1993).
Finally, we note that all of these models (Heit, 1998, 2000; Kemp &
Tenenbaum, 2009; Osherson et al., 2000; Sloman, 1993; Tenenbaum &
Griffiths, 2001) are challenged by Rotello and Heit’s (2009) finding that
logically valid arguments are judged to have less-plausible conclusions when
the arguments are longer. With the exception of Sloman’s model, each
model predicts that logically valid arguments are maximally strong. None of
these models predicts that adding premises to a logically valid argument will
weaken it. As noted by Lombrozo (2007) it is possible to develop a Bayesian
account of simplicity, but even the Bayesian models of inductive reasoning
predict that adding premises to a perfectly valid argument will not reduce
the probability of the conclusion.
OVERVIEW
In three experiments we investigated the effects of argument length on
judgements of plausibility. The baseline condition in Experiments 1a and 1b
served as replications of Rotello and Heit (2009). Both experiments also
included a new condition, forewarning, in which participants were
instructed to try to avoid using argument length in making plausibility
judgements. Experiment 1b used a somewhat stronger forewarning
manipulation than Experiment 1a, and also included an anti-forewarning
condition in which participants were encouraged to use argument length. In
general, because noticing argument length is automatic, because using
argument length is an intrinsic part of judging plausibility, and because
participants were not given any alternative means of making the
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
253
judgements, we predicted that it would be very difficult for participants to
control their own use of argument length. Experiment 2 added a withinparticipant manipulation of speaker reliability to the basic manipulation of
argument length. We expected that, overall, arguments from an unreliable
speaker would be rejected more than arguments from a reliable speaker. In
addition we examined whether use of logical information and superficial
information would vary for reliable versus unreliable speakers. One
possibility is that participants would attend to arguments from one type
of speaker more than the other, showing greater sensitivity to validity as well
as length. Another possibility, by analogy to Petty and Cacioppo (1984) is
that validity would matter more for one speaker and length would matter
more for the other speaker.
Although our focus was on whether the effects of argument length
changed as a function of instructions, to get a better understanding of the
results we applied the Rotello and Heit (2009) model to the data. Because
participants in each group were asked to make induction judgements we
expected that the same decision rule would apply, and therefore that the
slope of the decision bound would not differ greatly across conditions.
However, fitting the model could help explain other changes in results
across conditions. For example, potential differences in encoding
arguments could be manifested as different locations of the distributions
of the arguments.
EXPERIMENT 1A
This experiment, like Rotello and Heit (2009), examined the effects of length
and validity on judgements of inductive plausibility. The length manipulation involved varying the number of premises in the argument. The validity
manipulation involved presenting either invalid arguments or valid
arguments that maintained either an identity or class inclusion relation
between premise and conclusion categories. (See Osherson et al., 1990, for a
discussion of the theoretical importance of identity arguments; and see
Sloman, 1993, 1998, for further discussion of relations between identity and
inclusion arguments and the theoretical implications of studying both kinds
of arguments.) The key difference from Rotello and Heit was the
introduction of a forewarning condition.
Method
Participants. A total of 88 University of California, Merced undergraduate students, from a variety of majors, were paid to participate. They
were randomly assigned to one of two conditions: control (n ¼ 43) or
forewarning (n ¼ 45).
254
HEIT AND ROTELLO
Stimuli. There were 120 questions,1 comprising arguments about the
following kinds of mammals: bears, cats, cows, dogs, goats, horses, lions,
mice, rabbits, and sheep. An example invalid argument is:
Downloaded by [University of California Merced] at 08:37 11 June 2012
Horses have Property X
Mice have Property X
Sheep have Property X
——————————–
Cows have Property X
Note that we literally used ‘‘Property X’’. Participants were instructed to
treat this as a novel biological property. One-third of the arguments had a
single premise, that is, a single category above the line. One-third had three
premises (as in the previous example) and one-third had five premises. Half
the arguments were not deductively valid. The 60 remaining arguments were
deductively valid. Of these, 45 were identity arguments in which the
conclusion category was identical to one of the premise categories. An
example valid, identity argument is:
Horses have Property X
Mice have Property X
Sheep have Property X
Rabbits have Property X
Cats have Property X
———————————
Rabbits have Property X
The remaining 15 valid arguments were inclusion arguments in which the
conclusion category was included in one of the premise categories.2 The 3:1
ratio of identity versus inclusion relations was maintained for valid
arguments having one, three, or five premises. Following is an example of
a valid argument with an inclusion relation.
Mammals have Property X
————————————
Horses have Property X
1
We acknowledge that 120 questions is more than average for experiments on syllogisms.
Collecting so much data per participant has methodological advantages (e.g., providing
sufficient data for a complex design without the need for an unfeasible number of participants)
and potential disadvantages (e.g., diminishing quality of data over the course of an
experimental session). Although we did not address this latter issue systematically, performance
appeared to be similar earlier and later within each experiment. For example, in Experiment 1a,
d 0 (sensitivity to valid versus invalid arguments) was 1.7 for the first 60 trials and 1.8 for the last
60 trials.
2
Strictly speaking, inclusion arguments are enthymemes, because they rely on a hidden premise,
such as that all cows are mammals (Calvillo & Revlin, 2005). For simplicity we refer to both the
identity and inclusion arguments as valid.
ARGUMENT LENGTH AND INDUCTIVE REASONING
255
Downloaded by [University of California Merced] at 08:37 11 June 2012
Procedure. Each experiment was run using a program on a computer;
each participant took part individually. At the beginning of the experiment
the computer screen displayed instructions on the definition of strong
arguments. Specifically, following Rips (2001), participants were told that
strong arguments were those for which ‘‘assuming the information above
the line is true, this makes the sentence below the line plausible’’. In the
forewarning condition the following additional instructions were displayed:
You will see questions that vary in length. Sometimes you will only get one
sentence of information, but other times you will get several sentences of
information. Note that the length of argument is irrelevant to whether it is a
strong argument. Sometimes a short argument, with little information, is very
convincing, and sometimes a long argument, with a lot of information, is still
not very convincing at all. So in making your judgements about whether the
following arguments are strong, please try to IGNORE the length of the
question.
Following the instructions, 120 arguments were presented one at a time,
in a different random order for each participant. That is, different kinds of
arguments were intermixed, and participants were given the same
instructions for all of the arguments. Participants were told to assume that
the information above the line is true, and to assess whether the sentence
below the line was plausible. They pressed one of two keys to indicate
‘‘strong’’ or ‘‘not strong’’. Each binary decision was followed with a
confidence rating on a 1–5 scale; higher numbers indicated greater
confidence.
Results
Nine participants were excluded from the analyses because they gave the
same response for virtually every question, made more ‘‘strong’’ responses
to invalid than valid arguments, or had very low performance (d 0 .50
discriminating valid from invalid arguments). This eliminated three
participants from the forewarning condition and six from the control
condition.
Response rates. For an overview we first considered the probability that
participants endorsed the problem conclusion, as a function of validity,
number of premises, and condition. These data are shown in Table 1. For
both the control and forewarning conditions, increasing the length of invalid
arguments led to a higher rate of endorsement. It did not appear that the
forewarning instructions were effective in preventing the effect of argument
length. Looking at invalid arguments overall, participants in the forewarning condition were somewhat less likely to respond positively compared to
the control condition. Next, looking at valid arguments, there was a
Downloaded by [University of California Merced] at 08:37 11 June 2012
256
HEIT AND ROTELLO
tendency for longer arguments to be rejected more often than shorter
arguments, in both the control and forewarning conditions.
These results were subjected to separate ANOVAs for invalid and valid
arguments. For invalid arguments positive responses increased with the
number of premises, F(2, 154) ¼ 19.04, p 5.001, MSE ¼ .03, Z2 ¼ .19. The
effect of condition did not reach the level of statistical significance, F(1,
77) ¼ 2.52, p ¼ 0.12, MSE ¼ .22, Z2 ¼ .03, but condition interacted with the
number of premises, F(2, 154) ¼ 3.56, p 5.05, MSE ¼ .03, Z2 ¼ .04: The
effect of argument length was slightly stronger in the forewarning condition.
Each condition taken alone showed a significant effect of number of
premises on the responses to invalid arguments: control, F(2, 72) ¼ 6.02,
p 5.01, MSE ¼ .03, Cohen’s f ¼ .30; forewarned, F(2, 82) ¼ 17.18, p 5.001,
MSE ¼ .03; Cohen’s f ¼ .51.
For valid arguments positive responses decreased slightly with the
number of premises, F(2,154) ¼ 10.84, p 5.001, MSE ¼ .00, Z2 ¼ .12. There
was no effect of condition, F(1,77) ¼ 1.21, p ¼ 0.27, MSE ¼ .02, Z2 ¼ .02, and
no interaction of condition with number of premises, F(2,154) ¼ 1.43,
p ¼ 0.24, MSE ¼ .00, Z2 ¼ .02.
Signal detection analyses. Following the techniques used in previous
studies (Dube, Rotello, & Heit, 2010, 2011; Heit & Rotello, 2005, 2008,
2010; Rotello & Heit, 2009), we treated the reasoning task as a signal
detection task in which the goal was to discriminate strong arguments from
weak arguments. Simple accuracy measures such as the difference between
correct and error response rates, or the signal detection based d’, are often
confounded with participants’ overall tendency to accept a conclusion (see
TABLE 1
Response proportions from Experiments 1a, 1b, and 2
Invalid problems
Expt. 1a
Control
Forewarning
Expt. 1b
Control
Forewarning
Anti-forewarning
Expt. 2
Reliable
Unreliable
Valid problems
1 premise
3 premises
5 premises
1 premise
3 premises
5 premises
.35 (.06)
.18 (.04)
.36 (.05)
.32 (.04)
.48 (.05)
.39 (.05)
.98 (.00)
.97 (.01)
.96 (.01)
.97 (.02)
.95 (.01)
.93 (.02)
.09 (.02)
.07 (.02)
.07 (.02)
.14 (.03)
.14 (.04)
.11 (.02)
.20 (.04)
.15 (.04)
.24 (.05)
.93 (.01)
.89 (.03)
.93 (.01)
.87 (.02)
.85 (.02)
.84 (.02)
.86 (.02)
.83 (.02)
.85 (.02)
.24 (.04)
.09 (.03)
.38 (.05)
.17 (.03)
.54 (.05)
.26 (.05)
.95 (.01)
.80 (.04)
.85 (.03)
.75 (.04)
.89 (.02)
.77 (.03)
Standard errors are shown in parentheses.
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
257
Dube et al., 2010; Rotello, Masson, & Verde, 2008). To allow clear
evaluation of accuracy difference across conditions that are not confounded
with response bias differences we plotted receiver operating characteristic
(ROC) curves using the strong/not-strong responses as well as the 1–5
confidence ratings (for details see Macmillan & Creelman, 2005).
Figure 2 shows ROC curves for arguments of length 1, 3, and 5 in the
control condition (left panel) and forewarning condition (right panel). The
x-axis refers to the probability of responding ‘‘strong’’ on invalid items and
the y-axis refers to the probability of responding ‘‘strong’’ on valid items.
Each curve plots these probabilities for various levels of confidence
(response bias), with the left-most point on each curve reflecting the ‘‘sure
valid’’ response rates. Thus the points along each curve reflect the same
accuracy level but different response biases; more liberal response tendencies
increase both the correct and incorrect response rates yielding points
towards the upper-right in the space. Curves that fall higher in the space
(towards the upper-left corner) represent a greater degree of accuracy in
discriminating valid from invalid arguments because in that region the
correct response rate is high relative to the error rate. Accordingly, an
appropriate measure of performance is the area under the ROC, Az (Swets,
1986), which ranges from 0.5 (for chance performance) to 1.0 (perfect
accuracy).
It is evident that, in both conditions, longer arguments are associated
with a lower degree of accuracy. For five-premise arguments, participants
were most likely to respond positively to invalid arguments and most likely
to respond negatively to valid arguments. In comparison, for one-premise
Figure 2. Receiver operating characteristic (ROC) curves for Experiment 1b.
Downloaded by [University of California Merced] at 08:37 11 June 2012
258
HEIT AND ROTELLO
arguments, participants were less likely to respond positively to invalid
arguments and less likely to respond negatively to valid arguments. Hence
these ROC curves illustrate the effect of argument length. Indeed, the effect
of argument length appears stronger in the forewarning condition (right
panel), as there is more separation between the curve for length 1 and the
curves for length 3 and 5. For both the control and the forewarning
conditions there is greater area under the length 1 curve (Az ¼ .92 and .96,
respectively) than the length 5 curve (Az ¼ .86 and .87), reflecting greater
ability to discriminate strong from weak arguments for arguments with one
premise. These accuracy differences were significant in both conditions: For
the control group, z ¼ 7.10, p 5.001, using the methods described by Hanley
and McNeil (1983) for comparing dependent ROCs; in the forewarning
condition, z ¼ 12.57, p 5.001.
Modelling. We adopted Rotello and Heit’s (2009) approach to thinking
about these data, in terms of viewing argument strength as varying along
two underlying dimensions, associative strength (instantiated as argument
length) and apparent validity. Longer invalid arguments were allowed to
have greater strength on the associative dimension, but valid arguments with
more premises were allowed to have lower strength on that dimension. Valid
arguments have more strength on the dimension of validity. Because all
participants made induction judgements we assumed that the decision axis
was the same across conditions, reflecting the same relative weighting of the
two types of information. However, we did consider the possibility that
participants in the forewarning condition would pay more attention to
argument length during the encoding of the stimulus, which would imply
that the means of the distributions of argument strengths might differ across
conditions.
To evaluate the model we ran a large number of Monte Carlo simulations
in which the parameters were varied systematically. For each set of
parameter values we sampled 2000 simulated problem strengths from each
distribution; these strengths were compared to the decision bound, and
values falling in the ‘‘valid’’ response region were counted as resulting in a
‘‘valid’’ decision. This process yielded a simulated hit and false alarm rate
for each problem type for that set of parameter values; the simulated ROC
was mapped out by systematically varying the y-intercept of the decision
bound from high (yielding conservative responding) to low (resulting in
more liberal responding). The best-fitting set of parameter values was
chosen from those that produced both a relatively small mean-squared error
of prediction and a predicted area under the simulated ROC that fell within
the 95% confidence bands for the observed ROC. However, these
parameters should be considered illustrative only, as the entire parameter
space has not been searched exhaustively.
259
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
Estimated parameter values shown in Table 2. The only difference
between the control and forewarning conditions is the location of argument
distributions for one-premise valid problems, for control versus forewarning
condition, reflecting the somewhat higher level of accuracy on those items,
and the greater effect of argument length in the forewarning condition.
There is nothing in the estimated parameter values to indicate that
participants in the forewarning condition were less sensitive to argument
length than participants in the control condition.
The model predictions are shown in Figure 3, based on the estimated
parameters. The key aspects of the observed ROCs are captured in this
simulation. Most important, there is an effect of number of premises in both
conditions such that accuracy falls as number of premises increases, because
there are more positive responses to invalid arguments and fewer positive
responses to invalid arguments. The only difference between the two
conditions is that in the forewarning condition, the model accounts for the
TABLE 2
Parameter values for the two-dimensional model as applied to each experiment
Experiment
Parameter
dx ¼ mean of valid 1-premise
arguments on x-axis
Variance of dx
dy ¼ mean of valid 1-premise
arguments on y-axis
Variance of dy
Location of valid 3-, 5- premise
arguments on x-axis
Change in dy for valid 3-, 5-premise arguments
Location of invalid 3-, 5-premise
arguments on x-axis
Control or reliable condition slope
Forewarning or unreliable condition slope
Anti-forewarning condition slope
Change in dx for 1-premise valid arguments
in forewarning or unreliable condition
Change in dy for 1-premise valid arguments
in forewarning unreliable condition
Covariance of x and y for valid arguments
Covariance of x and y for invalid arguments
1a
1b
2
0.5
0.5
0.5
1
2.1
1
2.5
1
2.5
0.5
70.3, 70.3
0
0, 0.2
0.5
70.5, 70.5
0
0.1, 0.2
70.3
70.3
NA
0.2
70.3
70.3
70.3
0
0.7
0
0
0.2
0
0.2
2.0
71.0, 71.0
70.3
0.2, 0.4
70.3
70.5
NA
70.2
70.3
0.4
0.2
The distribution of invalid low-similarity arguments was located at (0, 0) with a variance of 1 on
each dimension and a small covariance.
Downloaded by [University of California Merced] at 08:37 11 June 2012
260
HEIT AND ROTELLO
Figure 3. Predicted ROCs and observed 95% confidence bands for data for Experiment 1a. (The
colour version of this figure is available in the online article.)
even higher level of accuracy on 1 premise items, which is part of the greater
effect of number of premises in the forewarning condition.
Discussion
The control condition of Experiment 1a replicated Rotello and Heit (2009)
in terms of showing that invalid arguments are considered stronger when
they are longer, and valid arguments are considered weaker when they are
longer. These results were shown using conventional analyses in terms of
response proportions as well as analyses in terms of area under ROC curves
plotted using confidence ratings.
The results of the forewarning condition were similar to the control
condition. If anything, there was a slightly stronger effect of argument
length in the forewarning condition. Clearly, warning participants not to use
argument length did not discourage them from doing so. It appears that
argument length is such a compelling attribute, both in terms of making
invalid arguments seem strong and in terms of making valid arguments seem
weak, that it is very difficult to ignore, even whether there is an explicit
request from the experimenter to do so.
In Experiment 1b we attempted a stronger forewarning manipulation,
and we examined the issue of whether the use of argument length can be
influenced in another way, by encouraging a group of participants to try to
show an argument length effect.
ARGUMENT LENGTH AND INDUCTIVE REASONING
261
Downloaded by [University of California Merced] at 08:37 11 June 2012
EXPERIMENT 1B
In this experiment we strengthened the forewarning manipulation somewhat
by presenting the instructions to ignore argument length twice, on paper
before the experiment as well as on the computer during the experiment. The
instructions were provided with the experimenter present. We also included
an anti-forewarning condition, in which participants were told that argument
length is important and were encouraged to consider argument length. This
experiment was patterned after a study by Heit, Brockdorff, and Lamberts
(2004), who found that a memory illusion (false alarms to semantic associates
of studied words) could not be reduced by a warning, but that participants
could be encouraged to exaggerate the illusion (make more false alarms to
semantic associates). In the anti-forewarning condition, if participants are
responding based on their inferences about the experimenter’s intent, they
should respond even more strongly to argument length.
Method
The method was the same as Experiment 1a except for the following: 135
individuals participated; control (n ¼ 45), forewarning (n ¼ 44), anti-forewarning (n ¼ 45).
The control condition was the same as in Experiment 1a. Participants in the
forewarning condition received the additional instructions, to ignore
argument length, twice. They first received these on a sheet of paper, before
the computer-based part of the experiment began. The experimenter watched
the participant read these instructions and asked whether he or she had any
questions. Once the computer-based experiment itself began, the forewarning
instructions were displayed on the computer screen as in Experiment 1a.
Participants in the anti-forewarning conditions also received additional
instructions, twice. However, they were instructed to consider argument
length when evaluating arguments. In particular, they were told:
You will see questions that vary in length. Sometimes you will only get one
sentence of information, but other times you will get several sentences of
information. Note that the length of argument is important to whether it is a
strong argument. Often a short argument, with little information, is not very
convincing at all, and often a long argument, with a lot of information, is very
convincing. So in making your judgements about whether the following
arguments are strong, please try to CONSIDER the length of the question.
Results
A total of 14 participants were excluded from the data analyses (4, 3, and 7
in the control, forewarning, and anti-forewarning conditions, respectively),
according to the same criteria as Experiment 1a.
Downloaded by [University of California Merced] at 08:37 11 June 2012
262
HEIT AND ROTELLO
Response rates. For an overview we first considered the probability that
participants endorsed the problem conclusion, as a function of validity,
number of premises, and condition. These data are shown in Table 1. For
the control, forewarning, and anti-forewarning conditions, increasing the
length of invalid arguments led to a higher rate of endorsement. It did not
appear that the forewarning instructions were effective in preventing the
effect of argument length—the results are close to the control condition.
Likewise, the anti-forewarning instructions did not seem to have much
effect. Next, looking at valid arguments, there was a tendency for longer
arguments to be rejected more often than shorter arguments, in all three
conditions. Overall the pattern of results is similar to Experiment 1a, except
that there is even less difference between the control and forewarning
conditions, and the overall rate of responding positively is somewhat lower.
For this latter result we do not offer an interpretation, but note that the
average rate of positive responding is not crucial to the issues investigated
here.
The results were subjected to ANOVA. On invalid arguments there was a
significant effect of number of premises, F(2, 234) ¼ 26.88, p 5.001,
MSE ¼ .02, Z2 ¼ .19, with more positive responses to longer arguments. The
effect of condition did not reach the level of statistical significance, F(2, 117)
51. The interaction of number of premises with condition did not quite
reach the level of significance, F(3, 184) ¼ 2.44, p ¼ .06, MSE ¼ .03, Z2 ¼ .04.
On valid arguments there was a significant effect of number of premises,
F(2, 234) ¼ 28.20, p 5.001, MSE ¼ .18, Z2 ¼ .19, with more positive
responses to shorter arguments. There was no effect of condition,
F(2, 117) 51, and no interaction of condition with number of premises,
F(2, 234) 51.
Signal detection analyses. Figure 4 shows ROC curves for arguments of
length 1, 3, and 5 in the control condition (left panel), the forewarning
condition (right panel), and the anti-forewarning condition (bottom panel).
As in Experiment 1a, in all conditions, longer arguments are associated with
a lower degree of accuracy: Those ROCs fall lower in the space. There are
more positive responses to longer invalid arguments, and fewer positive
responses to longer valid arguments. Given the visual similarity of the ROCs
produced in each of our three conditions, and the ANOVA results, we did
not expect to find differences in reasoning accuracy across conditions.
As in Experiment 1a we compared the length 1 versus length 5 curve in
each condition. For the control, forewarning, and anti-forewarning
conditions there was greater area under the length 1 curve than the length
5 curve, reflecting greater ability to discriminate strong from weak
arguments for arguments with one premise. Hence there were robust effects
of argument length in each condition. Using Hanley and McNeil’s (1983)
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
263
Figure 4. Receiver operating characteristic (ROC) curves for Experiment 1b. (The colour
version of this figure is available in the online article.)
method, the highly-significant z-statistics were 11.31, 9.43, and 11.38 for the
control, forewarning, and anti-forewarning conditions.
Modelling. Using the same strategy as in Experiment 1a we fitted
Rotello and Heit’s (2009) two-dimensional model to these data. The
resulting parameter values are shown in Table 2, and the simulated ROCs
are shown in Figure 5. Because the overall level of positive response was
lower in Experiment 1b, and there was less difference between the control
and forewarning conditions, some parameter values differed slightly from
Downloaded by [University of California Merced] at 08:37 11 June 2012
264
HEIT AND ROTELLO
Figure 5. Predicted ROCs and observed 95% confidence bands for Experiment 1b. Upper row:
control condition; middle row: forewarning condition; lower row: anti-forewarning condition.
(The colour version of this figure is available in the online article.)
Experiment 1a. Most important, exactly the same parameter values were
used to model the three conditions in Experiment 1b. Therefore the
modelling provides no evidence that participants’ encoding strategies
differed across conditions (the distributions have the same locations in all
conditions), and also no evidence that participants’ response strategies
differed across conditions (the same decision bound was used in all cases).
This modelling work suggests that participants in Experiment 1b were
unable or unwilling to avoid or exaggerate the influence of argument length,
even when asked.
Discussion
As in Experiment 1a there was an overall effect of argument length, with
invalid arguments seeming stronger as they get longer, and valid arguments
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
265
seeming weaker as they get longer. Despite the stronger, in-person, request
in this experiment, again the forewarning instructions were ineffective in
reducing the impact of argument length. Moreover, in the anti-forewarning
condition encouraging participants to use argument length also had little
effect, as if participants were already making maximum use of argument
length in the control condition. Unlike the Rotello and Heit (2009) study,
which showed robust effects of instructions (to judge validity or
plausibility), here participants’ judgements were not influenced by instructions. We conclude that argument length has intrinsic effects on inductive
reasoning.
Other than a difference in the average rate of positive responding, there
were minimal differences in the results of the two experiments and
accordingly in the estimated parameter values for the fitted model. None
of these differences suggests that experimenter instructions affected
participants’ use of argument length as a basis for judging plausibility.
For both experiments the model from Rotello and Heit (2009) gave a
satisfactory account of the detail results in terms of ROC curves.
Experiments 1a and 1b may have implications for another account of
inductive reasoning, relevance theory (Medin, Coley, Storms, & Hayes,
2003), which conceives of experiments on reasoning within a social context,
in which the experimenter is expected to follow Gricean principles of
informativeness and relevance (Grice, 1975). Simply put, making an
argument longer might reveal what the experimenter intends to convey
about the conclusion. Although Medin et al. did vary the length of
arguments, they did not explicitly address the possibility that length itself is
a cue to participants. If responding to the length of an argument reflects an
inference about the experimenter’s communicative intent, then an explicit,
even in-person, instruction from the experimenter to ignore length should at
least reduce the influence of length. Instead we found no influence of this
instruction. We would not minimise the importance of the social context of
the experiment, or Gricean inferences, when participants make judge
inductive plausibility, and we find Medin et al.’s (2003) relevance theory of
induction to be appealing in general (but see Heit & Feeney, 2005; Heit,
Hahn, & Feeney, 2005). However, in the present case the effects of argument
length seem to be due to the core nature of plausibility judgement, rather
than the participant’s perception of the experimenter’s communicative
intent.
EXPERIMENT 2
In this experiment we manipulated speaker characteristics: Each argument
was cued as coming from a reliable or an unreliable speaker. Overall we
Downloaded by [University of California Merced] at 08:37 11 June 2012
266
HEIT AND ROTELLO
expected that arguments coming from a reliable speaker would be
considered more plausible than arguments coming from an unreliable
speaker. This prediction simply assumes that participants were influenced by
the reliability cue in the direction of the cue. Another way of stating this
prediction is that reliability information is a kind of ad hominem comment
that is treated as a reasonable part of an argument (cf., Van Eemeren,
Garssen, & Meuffels, 2012 this issue).
In addition we examined how this manipulation interacted with
sensitivity to validity. It is possible that arguments from unreliable speakers
would receive greater scrutiny, leading to greater sensitivity to validity. Such
a result would be consistent with what is commonly reported as a belief bias
effect, namely that there is greater sensitivity to validity for unbelievable
conclusions than for believable conclusions (e.g., Thompson & Evans, 2012
this issue). However, Dube et al. (2010, 2011) showed that this commonly
reported result is a measurement artefact, and reported no true difference in
sensitivity. Another possibility is that arguments from unreliable speakers
are more likely to be rejected without consideration of their content. By this
account, participants would be less sensitive to validity for unreliable
speakers. Indeed, Hahn, Harris, and Corner (2009) found greater sensitivity
to evidence strength for reliable sources than for unreliable sources, in an
experiment on argument from ignorance. (For a review of other studies with
similar results see Pornpitakpan, 2004.) Another reason to make this same
prediction is in term of assimilation and contrast effects (Bohner, Ruder, &
Erb, 2002). Valid arguments coming from reliable speakers may seem
particularly strong, due to an assimilation effect, but invalid arguments from
reliable speakers may seem particularly weak due to a contrast effect. To the
extent that any such effects would be weaker for unreliable speakers, we
would expect greater sensitivity to validity for arguments coming from
reliable speakers.
Finally, with regard to sensitivity to argument length, a number of
predictions are possible. Given the findings from Experiments 1a and 1b
that participants have great difficulty not attending to argument length, one
would expect that sensitivity to validity and sensitivity to length would go
hand in hand. If arguments from unreliable speakers get more scrutiny, then
participants should be more sensitive to both validity and length for these
arguments. On the other hand, if arguments from unreliable speakers tend
to be dismissed without consideration of their content, then participants
should be more sensitive to both validity and length for arguments for
reliable speakers. Another possibility is that sensitivity to length is affected
differently from sensitivity to validity by speaker reliability. For example,
paying attention to relatively superficial cues such as argument length may
be more congruent with processing messages from unreliable speakers than
from reliable speakers. Relatedly, Petty and Cacioppo (1984) found that
ARGUMENT LENGTH AND INDUCTIVE REASONING
267
superficial features had a greater effect for low-involvement issues than highinvolvement issues.
Downloaded by [University of California Merced] at 08:37 11 June 2012
Method
The method was the same as the control condition of Experiments 1a and 1b
except for the following: 50 individuals participated. After the main set of
instructions, participants were given additional information:
Note that each argument comes from one of two sources, Jane or Mary.
Jane is a reliable source. 65% of her conclusions are strong. Mary is an
unreliable source. 35% of her conclusions are strong. When you judge
whether you think each conclusion is strong or weak, you should consider
the reliability of the source.
As each argument was presented for evaluation, at the top of the
computer screen it was noted either ‘‘Jane says (reliable, 65% strong)’’ or
‘‘Mary says (unreliable, 35% strong)’’.
Because speaker reliability was a within- participant manipulation we
increased the total number of arguments from 120 to 144, to increase the
amount of data per participant per condition. The arguments were
presented in a random order, intermixing arguments from the two speakers.
Half of the arguments were said to come from the reliable speaker and half
were said to come from the unreliable speaker. Note that this was a
completely uninformative cue: For each speaker 50% of arguments were
valid. In other respects the arguments maintained the same proportions as in
the other experiment. For valid arguments there was a 3:1 ratio of identity
to inclusion arguments. Overall, one third of the arguments had one
premise, one third had three premises, and one third had five premises.
Results
Six participants were excluded from the data analyses according to the same
criteria as Experiments 1a and 1b.
Response rates. For an overview we first considered the probability that
participants endorsed the problem conclusion, as a function of validity,
number of premises, and speaker reliability. These data are shown in Table
1. The manipulation clearly had a strong effect: Arguments from the reliable
speaker were more likely to be accepted than arguments from the unreliable
speaker. For both reliable and unreliable speakers, increasing the length of
invalid arguments led to a higher rate of endorsement. Next, looking at valid
arguments, there was a similar pattern for reliable and unreliable speakers:
Increasing argument length made three-premise arguments weaker than
one-premise arguments, but five-premise arguments were slightly stronger
Downloaded by [University of California Merced] at 08:37 11 June 2012
268
HEIT AND ROTELLO
than three-premise arguments. In general these patterns seem to replicate
Experiments 1a and 1b.
The results were subjected to ANOVA. On invalid arguments there was a
significant effect of number of premises, F(2, 84) ¼ 25.61, p 5.001,
MSE ¼ .02, Z2 ¼ .38, with more positive responses to longer arguments.
There was a main effect of speaker reliability, F(1, 42) ¼ 39.79, p 5.001,
MSE ¼ .074, Z2 ¼ .49, such that arguments from reliable speakers were
accepted more than arguments from unreliable speakers. The interaction
between number of premises and speaker reliability was also statistically
significant, F(2, 84) ¼ 6.21, p 5.01, MSE ¼ .014, Z2 ¼ .13. There was a
greater effect of number of premises for reliable speakers than for unreliable
speakers.
On valid arguments there was a significant effect of number of premises,
F(2, 84) ¼ 11.63, p 5.001, MSE ¼ .009, Z2 ¼ .22. There was a main effect of
speaker reliability, F(1, 42) ¼ 15.86, p 5.001, MSE ¼ .062, Z2 ¼ .27, such
that arguments from reliable speakers were accepted more than arguments
from unreliable speakers. The interaction between number of premises and
speaker reliability did not reach the level of statistical significance, F(2,
84) ¼ 1.48, p ¼ .23.
Signal detection analyses. Figure 6 shows ROC curves for arguments of
length 1, 3, and 5 for reliable (left panel), and unreliable (right panel)
speakers. As in Experiments 1a and 1b, in both conditions longer arguments
are associated with a lower degree of accuracy. There are more positive
Figure 6. Receiver operating characteristic (ROC) curves for Experiment 2.
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
269
responses to longer invalid arguments, and fewer positive responses to
longer valid arguments.
When comparing the curves for reliable versus unreliable speakers,
three trends are evident. First, it appears that participants were more
accurate at distinguishing valid from invalid arguments for reliable
speakers than for unreliable speakers, at least for arguments of length 1:
The ROC for reliable speakers falls higher in space than that for
unreliable speakers. We again used Hanley and McNeil’s (1983) method
to compare areas under these dependent curves. For length 1 arguments
the difference between reliable and unreliable speakers was significant,
z ¼ 6.47; for lengths 3 and 5 the differences were not significant (z ¼ 1.19
and 1.12, respectively).
Second, participants responded more liberally to the reliable sources; the
operating points are shifted to the right along the ROCs. We estimated ca
for each participant, assuming a zROC slope of 0.61 (obtained from the
group ROCs), and subjected those bias estimates to ANOVA. Participants
set more liberal criteria for longer arguments, F(2, 84) ¼ 11.54, p 5.001,
MSE ¼ .079, Z2 ¼ .22, and for the reliable source, F(1, 42) ¼ 46.12, p 5.001,
MSE ¼ .148, Z2 ¼ .52; the two factors did not interact F(2, 84) ¼ 1.80,
p ¼ .17.
The third trend was that it appears that participants’ responses were more
affected by argument length for reliable speakers than for unreliable
speakers. This result is evident from the separation of the curves. The spread
of curves from length 1 to length 5 is greater for reliable speakers than for
unreliable speakers. For both conditions, the difference between length 1
and length 5 is statistically significant (reliable, z ¼ 13.79; unreliable,
z ¼ 12.22).
Modelling. Again we fitted Rotello and Heit’s (2009) two-dimensional
model to these data. The resulting parameter values are shown in Table
2, and the simulated ROCs are shown in Figure 7. With regard to Figure
7, again the model predictions fall in a satisfactory way within the 95%
confidence intervals of the data. Crucially, parameter values differed by
condition (see Table 2). The locations of the valid distributions were
shifted down and to the left for the unreliable condition relative to the
valid condition. In other words, for the unreliable condition compared to
the reliable condition, valid items were closer to invalid items. This
finding corresponds to lower sensitivity to both validity (y axis) and
length (x axis) for the unreliable condition. In addition, the slope of the
decision boundary is somewhat steeper for the unreliable condition than
for the reliable condition, suggesting somewhat greater use of length
relative to validity, in the unreliable condition compared to the reliable
condition.
Downloaded by [University of California Merced] at 08:37 11 June 2012
270
HEIT AND ROTELLO
Figure 7. Predicted ROCs and observed 95% confidence bands for Experiment 2. Upper row:
reliable source condition; lower row: unreliable source condition. (The colour version of this
figure is available in the online article.)
Discussion
The analyses for Experiment 2 converge well. In comparison to Experiments
1a and 1b, Experiment 2 showed a more dramatic effect of the instructional
manipulation. Compared to reliable arguments, evaluation of arguments
from unreliable speakers was more likely to lead to rejection, less sensitive to
argument validity, and less sensitive to argument length. It appears that
participants were more engaged with arguments from the reliable speaker,
taking account of both validity and number of premises. Arguments from
the unreliable speaker were more likely to be rejected out of hand,
consequently taking less account of both validity and number of premises.
This finding is consistent with previous results from the argumentation and
social cognition literatures (Bohner et al., 2002; Hahn et al., 2009;
Pornpitakpan, 2004). This finding is the opposite of what is commonly,
but incorrectly, reported as a belief bias effect, that people are less sensitive
to validity for believable conclusions compared to unbelievable conclusions
(Dube et al., 2010, 2011).
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
271
Our modelling also suggested that, for the unreliable speaker, there was
proportionately greater use of more superficial information, related to
length, rather than validity. However, we interpret this result with caution
because there was less use overall of both validity and length in the
unreliable condition.
In other respects the results from Experiment 2 resembled those from
Experiments 1a and 1b. Most importantly, the same effects of argument
length appeared in both conditions. For invalid arguments longer
arguments were considered stronger. For valid arguments longer arguments
were generally weaker (comparing length 1 arguments to length 3 and 5
arguments). Hence the pervasive effect of argument length was pervasive
even across this manipulation.
Finally, we briefly turn to implications of our reliability manipulation for
the argumentation literature. As reviewed by Hahn and Oaksford (2012)
there have been multiple theoretical approaches to evaluation of argument
quality. One classic approach is the epistemic approach, being concerned
with matters of truth. Possibly listeners could infer that unreliable speakers
have poorer access to the truth. Bayesian approaches to argumentation (e.g.,
Hahn et al., 2009; Hahn & Oaksford, 2007) treat evaluation as a matter of
evidence and belief. Listeners’ prior beliefs could embody the assumption
that unreliable speakers’ arguments are less likely to be correct, and these
beliefs could be revised upward or downward depending on the content of
arguments. An alternative approach is the pragma-dialectical approach
(e.g., Rips, 1998; Van Eemeren et al., 2012 this issue; Van Eemeren &
Grootendorst, 2004), which is concerned with procedural aspects of
argumentation. Listeners could infer that unreliable speakers are less
competent at taking the steps necessary to make a sound argument, and this
inference could lead to more negative evaluations of arguments from
unreliable speakers. All of these accounts could readily explain the main
effect of reliability, but each account would need to be expressed in greater
detail to address the other results of Experiment 2, such as greater sensitivity
for both length and validity for arguments coming from reliable speakers.
GENERAL DISCUSSION
Using traditional analyses, ROC analyses, and modelling techniques, we
have converged on a consistent pattern of results. Replicating Rotello and
Heit (2009), when an argument is invalid, making it longer makes it seem
more plausible, and when an argument is valid, making it longer makes it
seem less plausible. The effects of argument length were so pervasive that
warning participants not to be influenced by length did not reduce its
influence. Participants were unwilling or unable to avoid the influence of
argument length, even when asked. Moreover, participants had so little
Downloaded by [University of California Merced] at 08:37 11 June 2012
272
HEIT AND ROTELLO
control over the influence of argument that they did not show greater effects
of argument length when encouraged to do so. Likewise, a manipulation of
speaker reliability had substantial effects on argument evaluation, but these
effects of argument length were robust for both reliable and unreliable
speakers.
Looking across several areas of reasoning research there have been
results pointing to increased strength with increased length (Burnstein et al.,
1973; Calder et al., 1974; Cook, 1969; Feeney, 2007; Gutheil & Gelman.,
1997; Lopez et al., 1992; Nisbett et al., 1983; O’Keefe, 1997, 1998, 1999;
Osherson et al., 1990; Petty & Cacioppo, 1984; Sloman, 1993) as well as
decreased strength with increased length (Lombrozo, 2007; Lopez et al.,
1992; Osherson et al., 1990; Petty & Cacioppo, 1984; Read & MarcusNewhall, 1993; Sloman, 1993). Our own contribution has been to document
both effects of argument length, within the same task, while holding most
things constant and varying only whether the argument is logically valid.
We do not presume to have developed a complete account of when
making arguments longer will make them stronger or weaker for different
reasoning tasks. However, in addressing this interesting issue, we do think it
will be valuable for future theoretical accounts of reasoning to cross task
boundaries, e.g., to explain results from argumentation, attitude change,
causal reasoning, and category-based induction paradigms, rather than
treating these as separate issues. Perhaps by treating each of these as signal
detection tasks, and applying multidimensional models of reasoning, some
progress can be made. However, our own signal detection models of
reasoning are not fully developed process models; we see these as potentially
constraining and informing process models of reasoning (for further
discussion see Dube et al., 2010, 2011; Heit & Rotello, 2010; Rotello &
Heit, 2009).
More generally, our results are compatible with recent theoretical work
on two-process accounts of reasoning (Evans, 2008; Stanovich, 2009):
participants are heavily influenced by automatically available information,
such as argument length, and can only override this information in limited
situations, for example when there is an alternative basis for response.
Because our participants were simply instructed to judge plausibility, there
was no readily available alternative.
Although our preference is for models that cut across experimental
paradigms, we conclude by discussing implications for models of categorybased induction, corresponding most closely to the paradigm in our own
experiments. As we have noted, most models of category-based induction
(Heit, 1998, 2000; Kemp & Tenenbaum, 2009; Osherson et al., 2000;
Sloman, 1993; Tenenbaum & Griffiths, 2001) assume that argument length
and plausibility judgement are closely connected: Each of these models
predicts robust effects of argument length on judgement. In fact, if we had
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
273
found that instructions were effective in reducing the effect of argument
length separately from the effect of validity, such results would have been
challenging for these models, which do not have a means for evaluating
plausibility that is not influenced by both length and validity.
The finding that valid arguments are considered weaker when they are
longer is somewhat problematic for these models. Although some of these
models (Osherson et al., 1990; Tenenbaum & Griffiths, 2001) predict
exceptions to the usual effect that longer arguments are stronger, these
exceptions involve changes in the superordinate category—for example,
from birds to animals in (3) and (4). That exception does not apply to our
experiments, in which the superordinate category was always mammal. Thus
these models predict that all of our longer arguments should be judged to be
stronger, yet participants consistently found the longer valid problems to be
weaker (replicating Rotello & Heit, 2009).
We see a few possible routes to augmenting models of category-based
induction. One possibility is to build in some notion of parsimony to
evaluation of arguments (Lombrozo, 2007). For example, if a model of
induction had an added component to give a slight penalty for longer
arguments, valid arguments could only lose plausibility with increased
length, because valid arguments would start with a very high level of
plausibility. On the other hand, invalid arguments might be able to
overcome the length penalty if making the argument longer increases its
plausibility by other means (e.g., in terms of the coverage component, in the
Osherson et al., 1990, model). Hence the model might predict a net increase
in plausibility as invalid arguments get longer, overcoming a slight penalty
for lack of parsimony. It may even be possible to build a more elaborate
Bayesian model taking such constraints into account (for some examples of
a related approach see Hahn et al., 2009; Hahn & Oaksford, 2007).
Another possible explanation is in terms of attention. For our valid
arguments, validity depended on one crucial premise: Arguments were valid
when one of the premises was a statement about mammals, or when there
was a match between a premise category and the conclusion category, e.g.,
they both concerned rabbits. Whereas judging that an invalid argument is
longer and hence more plausible does not depend on detailed scrutiny of
each premise, judging an argument to be valid does depend on identification
of the key premise. In a longer argument the crucial premise may be missed,
so longer valid arguments may be judged weaker, on average, than shorter
valid arguments. Comparing judgements about arguments from reliable
versus unreliable speakers, one key difference may be poorer attention to the
details of arguments from unreliable speakers. This attentional explanation
highlights some important asymmetries between invalid arguments and
valid arguments: Invalid arguments can vary greatly in their plausibility, but
valid arguments need to be perfect (Skyrms, 2000). And it is a useful
Downloaded by [University of California Merced] at 08:37 11 June 2012
274
HEIT AND ROTELLO
reminder that higher-level reasoning processes depend on lower-level
attentional processes.
Related to the both of the above explanations, the weakening of
conclusions when valid arguments are longer can be thought of as a kind of
dilution effect (e.g., Nisbett, Zukier, & Lemley, 1981; for the case that this
effect is normative see Tetlock, Lerner, & Boettger, 1996). Whatever
mechanisms are used to evaluate arguments could have limits on how much
information is processed, or could in effect average together all of the
evidence that is presented (cf., Anderson, 1965). Hence including nondiagnostic or weak evidence in addition to strong evidence could actually
make a conclusion seem weaker, if the weak evidence is replacing or being
averaged with the strong evidence.
To conclude, the pervasive effects of argument length on judgements of
plausibility, across multiple reasoning tasks, present some interesting
empirical puzzles and offer some challenges for future models of reasoning
that ideally will address how people weigh evidence in multiple reasoning
tasks.
REFERENCES
Anderson, N. H. (1965). Averaging versus adding as a stimulus-combination rule in impression
formation. Journal of Experimental Psychology, 70, 394–400.
Bohner, G., Ruder, M., & Erb, H-P. (2002). When expertise backfires: Contrast and
assimilation effects in persuasion. British Journal of Social Psychology, 41, 495–519.
Burnstein, E., Vinokur, A., & Trope. Y. (1973). Interpersonal comparison versus persuasive
argumentation: A more direct test of alternative explanations for group-induced shifts in
individual choice. Journal of Experimental Social Psychology, 9, 236–245.
Calder, B. J., Insko, C. A., & Yandell, B. (1974). The relation of cognitive and memorial
processes to persuasion in a simulated jury trial. Journal of Applied Social Psychology, 4, 62–
93.
Calvillo, D. P., & Revlin, R. (2005). The role of similarity in deductive categorical inference.
Psychonomic Bulletin & Review, 12, 938–944.
Cook, T. D. (1969). Competence, counter-arguing, and attitude change. Journal of Personality,
37, 342–358.
Dube, C., Rotello, C. M., & Heit, E. (2010). Assessing the belief bias effect with ROCs: It’s a
response bias effect. Psychological Review, 117, 831–863.
Dube, C., Rotello, C. M., & Heit, E. (2011). The belief bias effect is aptly named: A reply to
Klauer and Kellen (2011). Psychological Review, 118, 155–163.
Evans, J. St. B. T. (2008). Dual-processing accounts of reasoning, judgement and social
cognition. Annual Review of Psychology, 59, 255–278.
Feeney, A. (2007). How many processes underlie category-based induction? Effects of
conclusion specificity and cognitive ability. Memory & Cognition, 35, 1830–1839.
Grice, H. P. (1975). Logic and conversation. In P. Cole & J.L. Morgan (Eds.), Syntax and
semantics: Vol. 3. Speech acts (pp. 41–58). New York: Academic Press.
Gutheil, G., & Gelman, S. A. (1997). Children’s use of sample size and diversity information
within basic-level categories. Journal of Experimental Child Psychology, 64, 159–174.
Downloaded by [University of California Merced] at 08:37 11 June 2012
ARGUMENT LENGTH AND INDUCTIVE REASONING
275
Hahn, U., Harris, A. J. L., & Corner, A. J. (2009). Argument content and argument source: An
exploration. Informal Logic, 29, 337–367.
Hahn, U., & Oaksford, M. (2007). The rationality of informal argumentation: A Bayesian
approach to reasoning fallacies. Psychological Review, 114, 704–732.
Hahn, U., & Oaksford, M. (2012). Rational argument. In K. Holyoak & R. Morrison
(Eds.), The Oxford handbook of thinking and reasoning. Oxford, UK: Oxford University
Press.
Hanley, J. A., & McNeil, B. J. (1983). A method of comparing the areas under receiver
operating characteristic curves derived from the same cases. Radiology, 148, 839–843.
Hayes, B. K., Heit, E., & Swendsen, H. (2010). Inductive reasoning. Wiley Interdisciplinary
Reviews: Cognitive Science, 1, 278–292.
Heit, E. (1998). A Bayesian analysis of some forms of inductive reasoning. In M. Oaksford & N.
Chater (Eds.), Rational models of cognition (pp. 248–274). Oxford, UK: Oxford University
Press.
Heit, E. (2000). Properties of inductive reasoning. Psychonomic Bulletin & Review, 7, 569–
592.
Heit, E. (2008). Models of inductive reasoning. In R. Sun (Ed.), Cambridge handbook of
computational psychology. (pp. 322–338). Cambridge, UK: Cambridge University Press.
Heit, E., Brockdorff, N., & Lamberts, K. (2004). Strategic processes in false recognition
memory. Psychonomic Bulletin & Review, 11, 380–386.
Heit, E., & Feeney, A. (2005). Relations between premise similarity and inductive strength.
Psychonomic Bulletin and Review, 12, 340–344.
Heit, E., Hahn, U., & Feeney, A. (2005). Defending diversity. In W. Ahn, R. L. Goldstone,
B. C. Love, A. B. Markman, & P. Wolff (Eds.), Categorization inside and outside the
laboratory: Essays in honor of Douglas L. Medin (pp. 87–99). Washington, DC: American
Psychological Association.
Heit, E., & Rotello, C. M. (2005). Are there two kinds of reasoning? In Proceedings of the
Twenty-Seventh Annual Meeting of the Cognitive Science Society (pp. 923–928). Mahwah,
NJ: Lawrence Erlbaum Associates Inc.
Heit, E., & Rotello, C. M. (2008). Modelling two kinds of reasoning. In Proceedings of the
Thirtieth Annual Meeting of the Cognitive Science Society (pp. 1831–1836). Austin, TX:
Cognitive Science Society.
Heit, E., & Rotello, C. M. (2010). Relations between inductive reasoning and deductive
reasoning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36, 805–
812.
Heit, E., Rotello, C. M., & Hayes, B. K. (2012). Relations between memory and reasoning.
Psychology of learning and motivation, 57, 57–101.
Hornikx, J., & Hoeken, H. (2007). Cultural differences in the persuasiveness of evidence types
and evidence quality. Communication Monographs, 74, 443–463.
Kemp, C., & Tenenbaum, J. B. (2009). Structured statistical models of inductive reasoning.
Psychological Review, 116, 20–58.
Lombrozo, T. (2007). Simplicity and probability in causal explanation. Cognitive Psychology,
55, 232–257.
Lopez, A., Gelman, S. A., Gutheil, G., & Smith, E. E. (1992). The development of categorybased induction. Child Development, 63, 1070–1090.
Macmillan, N. A., & Creelman, C. D. (2005). Detection theory: A user’s guide (2nd ed.).
Mahwah, NJ: Lawrence Erlbaum Associates Inc.
McDonald, J., Samuels, M., & Rispoli, J. (1996). A hypothesis-assessment model of categorical
argument strength. Cognition, 59, 199–217.
Medin, D. L., Coley, J. D., Storms, G., & Hayes, B. K. (2003). A relevance theory of induction.
Psychonomic Bulletin & Review, 10, 517–532.
Downloaded by [University of California Merced] at 08:37 11 June 2012
276
HEIT AND ROTELLO
Nisbett, R. E., Krantz, D. H., Jepson, C., & Kunda, Z. (1983). The use of statistical heuristics in
everyday inductive reasoning. Psychological Review, 90, 339–363.
Nisbett, R. E., Zukier, H., & Lemley, R. (1981). The dilution effect: Nondiagnostic
information weakens the implications of diagnostic information. Cognitive Psychology,
13, 248–277.
O’Keefe, D. J. (1997). Standpoint explicitness and persuasive effect: A meta-analytic review of
the effects of varying conclusion articulation in persuasive messages. Argumentation &
Advocacy, 34, 1–12.
O’Keefe, D. J. (1998). Justification explicitness and persuasive effect: A meta-analytic review of
the effects of varying support articulation in persuasive messages. Argumentation &
Advocacy, 35, 61–75.
O’Keefe, D. J. (1999). How to handle opposing arguments in persuasive messages: A metaanalytic review of the effects of one-sided and two-sided messages. Communication
Yearbook, 22, 209–256.
Osherson, D. N., Smith, E. E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Petty, R. E., & Cacioppo, J. T. (1984). The effects of involvement on responses to argument
quantity and quality: Central and peripheral routes to persuasion. Journal of Personality and
Social Psychology, 46, 69–81.
Read, S. J., & Marcus-Newhall, A. R. (1993). Explanatory coherence in social explanations: A
parallel distributed processing account. Journal of Personality and Social Psychology, 65,
429–447.
Pornpitakpan, C. (2004). The persuasiveness of source credibility: A critical review of five
decades’ evidence. Journal of Applied Social Psychology, 34, 243–281.
Rips, L. J. (1998). Reasoning and conversation. Psychological Review, 105, 411–441.
Rips, L. J. (2001). Two kinds of reasoning. Psychological Science, 12, 129–134.
Rotello, C. M., & Heit, E. (2009). Modelling the effects of argument length and validity on
inductive and deductive reasoning. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 35, 1317–1330.
Rotello, C. M., Masson, M. E. J., & Verde, M. F. (2008). Type I error rates and
power analyses for single-point sensitivity measures. Perception & Psychophysics, 70, 389–
401.
Skyrms, B. (2000). Choice and chance: An introduction to inductive logic (4th ed.). Belmont, CA:
Wadsworth.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psychological Bulletin,
119, 3–22.
Sloman, S. A. (1998). Categorical inference is not a tree: The myth of inheritance hierarchies.
Cognitive Psychology, 35, 1–33.
Stanovich, K. E. (2009). What intelligence tests miss: The psychology of rational thought. New
Haven, CT: Yale.
Swets, J. A. (1986). Indices of discrimination or diagnostic accuracy: Their ROCs and implied
models. Psychological Bulletin, 99, 100–117.
Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalisation, similarity, and Bayesian inference.
Behavioral and Brain Sciences, 24, 629–641.
Tetlock, P. E., Lerner, J., & Boettger, R. (1996). The dilution effect: Judgemental bias,
conversational convention, or a bit of both? European Journal of Social Psychology, 26, 915–
935.
Thagard, P. (1989). Explanatory coherence. Behavioral and Brain Sciences, 12, 435–467.
Thompson, V., & Evans, J. St. B. T. (2012 this issue). Belief bias in informal reasoning. Thinking &
Reasoning, 18, 278–310.
ARGUMENT LENGTH AND INDUCTIVE REASONING
277
Downloaded by [University of California Merced] at 08:37 11 June 2012
Van Eemeren, F. H., Garsen, B., & Meuffels, B. (2012 this issue). The disguised abusive ad
hominem empirically investigated strategic manoeuvring with direct personal attacks.
Thinking & Reasoning, 18, 344–364.
Van Eemeren, F. H., & Grootendorst, R. (2004). A systematic theory of argumentation. The
pragma-dialectical approach. Cambridge, UK: Cambridge University Press.